

Post-Train Frontier Open Models at Scale
Your RL run can finish cleanly and still hand you a worse model.
A small mismatch between your trainer and inference engine can skew the training signal. Even with a clean signal, slow rollouts limit how many experiments you can run and raise the cost of each one. And once you find what works, results can drop when adapters are compressed for serving. All three get harder as you post-train frontier open models at scale.
In this session, our research team will dive into our new service, Together Post-Training, and how it handles these three problems. You'll see it firsthand as the team walks you through an RL run step by step.
What you’ll learn:
How to see the trainer-inference mismatch at every step, and how off-policy correction, on by default, accounts for it
How to run more experiments on the same GPUs, with Together’s serving optimizations tuned for each model and GPU type, plus several LoRA runs per reservation
How to keep the quality you trained when adapters are merged and served in FP4
How to post-train a Kimi K3.1 model with RL
Register to join the webinar live, ask questions, and receive a recording of the talk.
⚠️ Please provide your valid corporate email upon registration.